Embedded Files
Connecting your micro:bit
Connecting your micro:bit
Get started by connecting your micro:bit to your computer.
The micro:bit should already have the Jacdac software loaded.
On the Jacdac website (see the embedded page to the right), click on the Jacdac logo in the upper right corner to go to the Device Dashboard.
In the device dashboard select the 'Connect USB' button . This will bring up a dialog with a lists of connect micro:bit devices that you can pair with your computer.
Choose the device that you would like to work with and select 'Connect'. You will then see all of the services (e.g. buttons, accelerometers, etc.) that are associated with you device appear in the Dashboard.
Connecting your device to your computer.
Connecting your device to the web page.
Collecting Data
Collecting Data
Once your device is connected, use the browser back button to return to the ML Model Creator page.
At first, you will see that your current dataset is empty. You should also see all of the sensors connected to your micro:bit available as input sensors for your dataset.
Select the input sensor associated with accelerometer data, then a trend line will appear on the screen showing the reading from the accelerometer sensors.
The parameters in the Record data section allow you to create a dataset using that input sensor data. You can type a name for your recording in the Recording name box, you can type or select a label for the recording in the Class label box, and you can change the sample duration and interval with the last two boxes.
Key point: In this tutorial, you only need to worry about the class label. As you collect data for different samples, make sure that the class label matches up.
Try pressing the 'Start' button. After 2 seconds, a newly recorded sample will pop up in the Current Dataset section. Press the trashcan next to the recording to remove it.
Collecting accelerometer data with the web interface. You can collect data from any combination of input sensors.
In the Current Dataset section, use the trash can next to individual recordings to remove them. Or use the trashcan at the top of the section to remove all recordings.
Setting Labels
Setting Labels
Which labels. In the define section, we identified "Easy and natural" as a value for the kinds of gestures we want to classify. Gestures like taps, swipes, and hand rotations are all simple gestures most people can perform without needing to learn how.
Example gestures include taps, double taps, swipes, rotations, and much much more!
The parameters in the Record data section allow you to create a dataset using that input sensor data. You can type a name for your recording in the Recording name box, you can type or select a label for the recording in the Class label box, and you can change the sample duration and interval with the last two boxes.
Key point: In this tutorial, you only need to worry about the class label. As you collect data for different samples, make sure that the class label matches up.
Try pressing the 'Start' button. After 2 seconds, a newly recorded sample will pop up in the Current Dataset section. Press the trashcan next to the recording to remove it.
Use the 'Class label' box to indicate which class a recording belongs to. You can type in new classes or select with the dropdown list.
Case study: Representation in datasets
Case study: Representation in datasets
Making gesture interfaces accessible is a key area of research in machine learning. Some approach the problem by trying to collect datasets that are as large and broadly representative as possible. Another option is to collect very small, personalized datasets that are unique to each user.
iPhone Assistive Touch is an excellent example of how the iPhone is made more accessible by allowing users to create small, personalized gesture datasets. If accessibility is a priority in your project, and it probably should be, then consider if this separate approach is possible.
Data Quality Check
Data Quality Check
Effective datasets. We identified "accurate predictions" as an important value. There is some tension between having a very accurate systems and adding more classes of gestures. Recognizing more gestures increases the flexibility of the system, but may compromise the accuracy of classification. Let's stick with our four classes for now.
Key point: Machine learning can often classify things that human eyes cannot. Even so, doing a "common sense" comparison between classes will help us plan future evaluations of the model.
Representative datasets. In order for our model to work fairly well for lots of different people, we will need multiple examples of people performing these gestures in different ways. A special subgroup in our list of stakeholders are people who may perform gestures differently than others due to physical impairments or other factors.
Key point: We need a plan to collect data that ensures that we get sufficient and diverse samples, especially from subgroups.
Balanced datasets. If at all possible, we should strive to keep our dataset balanced, meaning that we have similar numbers of samples for every class. However, this is not always possible or easy - it's way easier to collect lots of different examples of hands being still in various orientations.
Key Point: Make a note of classes that need way more examples to properly represent their class. Make a note of classes that are very uniform and may have fewer classes.
Examine your dataset. Look for
- Classes that seem very similar to one another
- Samples that look like outliers in their class - you may need to collect more examples like it.
- Classes that have many more or fewer samples than others
The feature plot that displays beneath the dataset can help you find outliers and very similar classes
Other data collection considerations
Other data collection considerations
Informed consent. It is important to make sure that participants and stakeholders are informed if we are collecting data about them or about their environment. And, if at all possible, we should strive to avoid collecting unnecessary, sensitive information.
Anticipating unintended uses. This dataset only contains data about hand movements. Still, there may be unintended consequences of collecting and sharing this dataset. For example, if we do a good job of collecting data from people with a particular movement disorder, that protected health information might inadvertently be perceptible in our dataset. Someone could then misuse our dataset to detect if someone has that disorder based on how they perform gestures. We should take steps to ensure the data is only used in ways that we received consent for.
Putting it all together
Putting it all together
As you complete your dataset, take time to complete the following form about its attributes. This will be used to create a dataset nutrition label that will be packaged with your dataset CSV file when you export it.
Page updated
Google Sites
Report abuse